Author: Jiu Lin
Recently, Hygon, Cambricon, Moore Threads, Muxi Semiconductor, Biren Technology, and Tianshu Zhixin have sequentially disclosed their 2026 semi-annual reports. Enflame Technology, which was recently listed, also updated its data for the same period in its prospectus. The combined revenue of the seven companies in the first half of the year reached approximately CNY 21.46 billion, nearly double that of the same period last year.
Revenue has surged across the board: Cambricon grew by 108.13%, Moore Threads by 147.42%, Tianshu Zhixin by 191.6%, Enflame Technology by approximately 279%, and Biren Technology surged by 1997.6% on a low base. Domestic AI chip companies, which once relied on financing to support R&D, crossing the threshold from design verification to scaled mass delivery.
01 Who Makes the Most Money? Look at the Books First
"Who makes the most money" is not a question that can be answered simply by looking at the net profit ranking. Hygon has the highest revenue, while Cambricon has the highest net profit attributable to shareholders; Muxi and Tianshu Zhixin have already turned a profit on their statements, but their core businesses remain unprofitable; Moore Threads is just one step away from breaking even; Biren and Enflame have seen rapid revenue growth but have yet to shake off hundreds of millions in losses.
There are four relatively intuitive trends:
First, collective high growth in revenue, with multiple companies doubling. Cambricon's revenue grew by 108.13%, Moore Threads by 147.42%, Tianshu Zhixin by 191.6%, Enflame by approximately 279%, and Biren surged by 1997.6% on a low base. Although Hygon and Muxi did not double, their growth rates of 66.52% and 44.67% are also significantly higher than most mature chip companies. More importantly, the half-year revenue of several companies has already exceeded their full-year revenue last year. Moore Threads' first-half revenue was CNY 1.736 billion, higher than its full-year CNY 1.506 billion in 2025; Biren's first-half revenue was CNY 1.236 billion, also exceeding last year's full year; Enflame's half-year revenue of CNY 1.120 billion surpassed last year's full-year CNY 990 million. The business scale of domestic AI chip manufacturers is stepping onto a new level, indicating that domestic AI chips are moving from small-batch testing and project validation to larger-scale product and cluster delivery.
Second, the profitable camp and the cash-burning camp are beginning to diverge. While revenues are growing together, profitability has not converged simultaneously. Hygon and Cambricon have entered the stage of scaled core business profitability. Moore Threads and Muxi are in the second position, with both companies touching the break-even line for their core businesses. Moore Threads' net loss attributable to shareholders narrowed to CNY 11.56 million, and Muxi's non-recurring net loss attributable to shareholders for the half-year was CNY 49 million. Biren Technology's gross profit in the first half was CNY 527 million, yet it invested CNY 804 million in R&D; Enflame Technology's net loss was CNY 630 million, and it remains in the stage of simultaneous expansion of products and customers.
Third, gross margins remain resilient overall, with differences stemming more from product mix. Following the volume ramp-up of domestic AI chips, there has been no widespread descent into low-price market grabbing. Cambricon's overall gross margin in the first half was 55.25%, further rising to 56.10% in the second quarter; Muxi's gross margin was approximately 57.22%, a slight year-on-year increase of about 1.1 percentage points; Moore Threads' gross margin was 56.95%. The three companies still maintain margins above 55%, indicating that domestic chips retain a certain degree of pricing power in a market with tight supply and demand. Tianshu Zhixin's overall gross margin dropped from approximately 50.1% in the same period last year to 17.2%.
Fourth, R&D investment enters the realization period, with differences in conversion speed. High R&D investment is the norm in the chip industry. Looking at several GPU companies still ramping up: Moore Threads, Muxi, Biren, Tianshu Zhixin, and Enflame invested CNY 769 million, CNY 525 million, CNY 804 million, CNY 559 million, and CNY 640 million in R&D in the first half, respectively, totaling approximately CNY 3.297 billion, which accounts for about 52% of the combined revenue of the five companies in the same period. In contrast, Cambricon's R&D expenses in the first half were approximately CNY 703 million, on a similar scale to Moore Threads and Biren. Cambricon's profit performance further indicates that its existing products have entered the stage of scaled delivery and profit realization.
Expense ratios better reflect the stage differences among the companies. The R&D expense ratios of Moore Threads and Muxi are 44.3% and 39.65%, respectively, while those of Biren, Tianshu Zhixin, and Enflame are approximately 65%, 59.1%, and 57.1%. High R&D spending is merely a ticket to enter the game; what truly determines financial performance is whether the investment can form mass-producible products, whether the products can enter scaled clusters, and whether they can bring continuous orders. The competition among domestic AI chips is shifting from "daring to invest" to "ability to deliver".
02 Orders Booked Until Next Year: Domestic AI Chips in Short Supply
According to CCTV Finance, domestic computing power chips are experiencing an "explosion in orders," with the delivery schedule for some products already extended to a year later, and domestic computing power demand reaching more than 10 times the supply. Multiple computing power service providers stated that orders have maintained high-speed growth for two consecutive years, and the number of customers for some enterprises has increased by about 10 times over the past year. Demand is transmitting upstream along the industry chain, leading to tightness in chips, servers, storage, and packaging.
Behind the tight supply is the fact that domestic chips have begun to undertake real high-traffic services. In late August, Zhipu disclosed the deployment details of GLM-5.3-Flash, the first native multimodal model in the GLM-5 series: all online traffic is supported by 100,000 domestic chips. Before the official release of the model, Zhipu also tested it anonymously on two major overseas platforms, OpenRouter and OpenCode, with a cumulative token invocation volume reaching 62T, and all related request traffic was powered by domestic chips. After underlying architecture transformation and inference service optimization, the cluster hardware efficiency and single-token cost have reached a level comparable to mainstream NVIDIA GPUs. "This proves that domestic chips can completely and efficiently support the inference needs of frontier models in large-scale scenarios in an economical manner," Zhipu wrote in the announcement.
Such cases test not only peak computing power but also concurrent scheduling, inter-chip communication, fault recovery, model adaptation, and cost control. Therefore, customers' procurement standards have extended from the performance of a single card to the availability of the entire system. Domestic chips have moved from "being able to light up" to "providing stable services," and the supply-demand relationship has also begun to show new anchor points.
The demand structure is changing simultaneously. Training clusters remain the source of large-scale procurement, while inference, with the continuous release of search, recommendation, intelligent customer service, content generation, and enterprise AI agents, is becoming a more stable and higher-frequency consumption of computing power. Tianshu Zhixin's revenue structure has already shown this curve. In the first half of 2026, the company's inference series revenue reached CNY 654 million, a year-on-year growth of 651.8%, exceeding the training series' CNY 262 million and approaching twice the full-year inference revenue of 2025. The company stated that the increased sales volume and selling prices of high-end products jointly drove the growth of the inference business. For domestic chip companies, the inference market has opened a second entry point beyond training.
03 Commercialization Moves from Selling Cards to Delivering Systems
Based on the semi-annual reports, the commercialization of domestic AI chips has formed a clear systematic path: cloud products provide standardized revenue, cluster delivery scales up single orders, industry ecosystems expand the customer radius, and software adaptation and continuous services determine the depth of repurchases.
Cloud products are the starting point for revenue scaling. Cambricon's cloud product line revenue in the first half was CNY 5.994 billion, accounting for 99.98% of total revenue, with training and inference chips focusing on serving large model infrastructure. Hygon's DCU has completed the adaptation of over 400 mainstream large models, entering cloud service providers and intelligent computing centers relying on the synergy of CPUs and DCUs.
Moore Threads has further extended its platformization to clusters. The company's cloud product revenue in the first half was approximately CNY 1.69 billion, accounting for 97.49% of total revenue, with the mass production of MTT S5000 and the delivery of the Kuae intelligent computing cluster becoming the main drivers of growth. The CNY 660 million Kuae intelligent computing cluster sales contract disclosed in March has been fully delivered and recognized as revenue in the first half; the Kuae cluster has also been deployed in Beijing, Wuxi, Hangzhou, and other places. On September 9, JD Cloud announced the creation of a domestic 10,000-card cluster with partners including Moore Threads, and plans to build a 100,000-card full-function GPU cluster, extending the system delivery of domestic GPUs upward from 10,000-card projects.
Muxi proposed the "1+6+X" industrial ecosystem strategy, focusing on its independent GPU computing power platform and MXMACA software stack, covering six industries: finance, healthcare, energy, education and scientific research, transportation, and large entertainment, and extending to scenarios such as embodied intelligence and low-altitude economy. The domestically produced Xiyun C600 achieved mass production in May, and the Xijing S600 supernode for training, inference, and intelligent computing center construction is advancing simultaneously. Muxi GPUs have achieved commercial applications in thousand-card scale clusters.
Biren Technology follows the route of cloud training combined with high-speed interconnection. The R100 series targets intelligent computing centers, operators, and large model customers, with first-half revenue of CNY 1.236 billion, and the volume ramp-up of training products driving the gross margin up to 42.7%. Jointly released with Shanghai INESA, Lightelligence, and ZTE, the commercial version of the "Guangyue" optical interconnect and optical switching GPU supernode has achieved deployment on a scale of thousands of cards. In July, the next-generation optical interconnect supernode solution was announced, extending the product form from training chips to large-scale interconnection systems.
Enflame Technology serves internet customers such as Tencent with training and inference chips and cluster solutions, and the application of its products at Tencent has expanded from small-batch pilots to multiple scenarios. Tencent is not only an important shareholder of Enflame Technology but also its largest customer. In 2025, direct sales and related model revenue from Tencent accounted for 83.79%; beyond Tencent, some non-associated top internet customers have completed hardware and model testing and have begun to place small-batch orders.
04 Conclusion
The current high prosperity has given domestic AI chips a window for collective volume ramp-up. When supply is tight, customers' first priority is to get the goods; when capacity gradually recovers and chip prices and computing power service prices are renegotiated, customers will shift their comparison dimensions to utilization rates, energy consumption, software upgrades, and payment collection cycles.
The Ministry of Industry and Information Technology issued the "15th Five-Year Plan for the Development of the Information and Communications Industry," proposing to elevate the development level of computing power facilities, deepen the coordinated development capability of computing power, and strengthen public computing power service capabilities. By 2030, the scale of China's intelligent computing power will grow by more than 5 times compared to 2025. Orderly deploy 10,000-card, 100,000-card, and above intelligent computing clusters, deploy inference computing power facilities on demand for specific scenarios, and increase efforts to adapt domestic computing power chips.
Domestic AI chips are moving from "selling a single card" to "delivering a system, serving an industry, and operating an ecosystem." Going public and raising funds only buys time; the next round of rankings will not only look at who has the fastest revenue growth. Companies that navigate the computing power cycle need to be able to replicate one-off projects into clusters and settle these clusters into stable industry workloads.